A data-driven approach to refine predictions of differentiated thyroid cancer outcomes: a prospective multicenter study

Background: The risk stratification of patients with differentiated thyroid cancer (DTC) is crucial in clinical decision making. The most widely accepted method to assess risk of recurrent/persistent disease is described in the 2015 American Thyroid Association (ATA) guidelines. However, recent research has focused on the inclusion of novel features or questioned the relevance of currently included features. Objective: To develop a comprehensive data-driven model to predict persistent/recurrent disease that is able to capture all available features and determine the weight of predictors. Design: Prospective cohort study using the Italian Thyroid Cancer Observatory (ITCO) database (NCT04031339). Setting: Forty Italian clinical centers. Methods: We selected consecutive cases with DTC and at least early follow-up data (n=4773; median follow-up 26 months, interquartile range 12-46 months). A decision tree was built to assign a risk index to each patient. The model allowed us to investigate the impact of different variables in risk prediction. Results: 2492 patients (52.2%) were classified as low, 1873 (39.2%) as intermediate, and 408 as high risk, according to ATA risk estimation. The decision-tree model outperformed the ATA risk stratification system: the sensitivity of high-risk classification for structural disease increased from 37% to 49%, and the negative predictive value for low-risk patients increased by 3%. Feature importance was estimated. Several variables not included in the ATA system significantly impacted the prediction of disease persistence/recurrence: age, body mass index, tumor size, sex, family history of thyroid cancer, surgical approach, pre-surgical cytology, and circumstances of the diagnosis. Conclusions: Current risk stratification systems may be complemented by the inclusion of other variables in order to improve the prediction of treatment response. A complete dataset allows for more precise patient clustering.